[R] ROC for threshold value, biometrics

Frank E Harrell Jr f.harrell at vanderbilt.edu
Fri Jun 11 17:06:10 CEST 2004

myint.tin-tin-htar at jvr.ap-hop-paris.fr wrote:
> Hello,
> I am just a beginner of R 1.9.0.
> I try to construct a predictive score for the development of liver
> cancer in cirrhotic patients. So dependant variable is binanry (cancer
> yes or no). Independant variables are biological data. The aim is to
> find out a cut-off value which differentiate (theoratically)  from
> normal to pathological state for each biological data.

A binary endpoing is not a good choice for this problem, as the time to 
diagnosis is very important.  Unless you only have a biopsy at a single 
fixed time (e.g., 5 years post study entry) it would be good to consider 
for example a Cox proportional hazards model.  And there are many 
reasons for not using a cutoff, as detailed in my book Regression 
Modeling Strategies.

> How can I step in procedue to get a cut-off value (threshold) for each
> variable? I think I should try by ROC. But I'm not sure. If so, someone
> can lead me?

Besides not recommending the use of cutoffs on an overall predicted 
value, I think that using cutoffs based on separate analyses of 
predictors is even worse.


> If not, someone can advice me how ?
> Any advice will be cordially aprreciated.
> Tin Tin Htar Myint
> Research assistant
> Liver unit
> Jean Verdier Hospital, Bondy
> France
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://www.stat.math.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html

Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University

More information about the R-help mailing list